m4so/jaspeech

Japanese Anime Speech Dataset

日本語はこちら

japanese-anime-speech is an audio-text dataset designed for the training of automatic speech recognition models. The dataset is comprised of thousands of audio clips and their corresponding transcriptions from different visual novels.

The goal of this dataset is to increase the accuracy of automatic speech recognition models, such as OpenAI's Whisper, in accurately transcribing dialogue from anime and other similar Japanese media. This genre is characterized by unique linguistic features and speech patterns that diverge from conventional Japanese speech.

A list of all audio files and transcriptions are here.

Content Warning: Please be advised that the majority of the audio in this dataset is sourced from visual novels and may include content that is not suitable for all audiences, such as suggestive sounds or mature topics. Efforts have been undertaken to minimise this content as much as possible.

Dataset information

• 73,004 audio-text pairs
• 110 hours of audio (OpenAI suggests a minimum of 5 hours for productive Whisper fine-tuning).
• 5.4s average audio length
• Transcriptions have been scraped directly from the game files of visual novels
• Lastest version: V5 - March 22nd 2024

Changelog

• V1 - This version contains 16,143 audio-text pairs from the visual novel IxSHE Tell. Some cleaning of the transcriptions has been done to get rid of unwanted characters at the start and end of lines.
• V2 - The version contains 23,422 audio-text pairs from three different visual novels. Cleaning has been done to remove most nsfw lines, especially noises that aren't words. The audio is now in mp3 format, rather than wav. This version contains 32.6 hours of audio.
• V3 - The version contains 38,325 audio-text pairs from five different visual novels. Thorough cleaning has been done to remove most nsfw or low-quality audio files. Transcriptions have been formatted to contain much fewer dramatised duplicated characters (for example 「ああああーーー」), and transcriptions have been made much more consistent. This version contains 52.5 hours of audio.
• V4 - The dataset contains 47,844 audio-text pairs from six different visual novels. Thorough cleaning has been done to remove most nsfw or low-quality audio files. This version contains 63.4 hours of audio.
• V5 - The dataset contains 73,004 audio-text pairs from eight different visual novels. Thorough cleaning has been done to remove most nsfw or low-quality audio files. This version contains 110 hours of audio.

Bias and Limitations

This dataset, while valuable for training anime-style Japanese speech recognition, has some inherent biases and limitations. The audio is primarily sourced from visual novels, leading to a gender bias towards female voices and a domain-specific vocabulary revolving around topics such as love, relationships, and fantasy. Additionally, the professionally produced nature of the audio results in clear and slow speech, which may not fully reflect real-world speaking patterns.

Use & Credit

This dataset is openly available for commercial or non-commercial use. Anyone is welcome to use this dataset as they deem appropriate. However, the creator assumes no responsibility for the consequences of its use. While not mandatory, crediting this dataset with a hyperlink in any derivative work would be greatly appreciated.

I hope that by sharing this dataset, we (the open-source community) improve automatic speech recognition for anime content.